Back to PLS HelpsMRI Design Matrix query for Behavioural PLS
Posted on 07/01/10 12:45:00
Number of posts: 18
I'm worried about colinearity in my data. At the moment I have 2 groups with one sMRI scan per subject and one behavioural measure for each subject. I Ioad up my imaging data and my behavioural measure and run a behavioural PLS with one column in the text file with a number at each row corresponding to each individual's behav score. At the end I get a LV brain score that correlates with the behave score and a nice voxel-wise Brain Pattern associated with that LV. I then test the brain score for a group difference and I get a significant result. But what if there is already a group difference in my behavioural measure (which there is)? Or alternatively there are already group differences in rGM volume? Is there a confound?
thanks
AB
Replies:
Untitled Post
Posted on 07/02/10 08:48:24
Number of posts: 394
Hi Anna.
If I understand your analysis correctly, it would not be surprising to get a group difference in brain scores if there are differences in between the groups on regional gray matter volumes. Since the brain score is a weighted sum of all the voxels in an image, if there are global differences in voxel values between groups, these are likely to exist in the brain scores as well.
If your question is whether the groups differ in the relation between your behavior measure and rGM, I would do a two group behavior PLS. If there are group differences in that correlation, it will come out as a significant LV. In that case, group mean differences in behavior or brain will not be a confound because you are looking a variation around each group's mean.
Randy
If I understand your analysis correctly, it would not be surprising to get a group difference in brain scores if there are differences in between the groups on regional gray matter volumes. Since the brain score is a weighted sum of all the voxels in an image, if there are global differences in voxel values between groups, these are likely to exist in the brain scores as well.
If your question is whether the groups differ in the relation between your behavior measure and rGM, I would do a two group behavior PLS. If there are group differences in that correlation, it will come out as a significant LV. In that case, group mean differences in behavior or brain will not be a confound because you are looking a variation around each group's mean.
Randy
Untitled Post
Posted on 07/02/10 11:34:40
Number of posts: 18
Thanks Randy, that makes sense.
So for a 2 group Behaviour PLS do I need another column that has a list of 1 and 2 for group 1 and group 2.
thanks
Anna
Untitled Post
Posted on 07/05/10 15:51:41
Number of posts: 394
You'll need a single vector for the behavior, but make sure that you have one group stacked on top of the other rather than intermixed.
Untitled Post
Posted on 07/30/10 04:15:31
Number of posts: 18
Dear Randy
Just to clarify this point about group differences.
I followed your instructions re stacking the groups: so my session file had 1 condition and 100 subjects all GMV maps the controls are numbered 1-50 and the patients are numbered 51-100
I chose Behav PLS to run and entered the behav score in the same order as the images were uploaded; patients stacked on top of controls.
After 1000 permutations I get a very nice LV1 image that has an r value of ~0.9 of brain scores against behav measure. If I then do a t-test between the brain scores of group1 vs group2 I get a significant difference.
The image therefore is the pattern that best explains the variance associated with the beahv score in both groups but the blue areas are decreases in GMV in controls compared to patients and the yellow areas are increases in GMV compared to patients.
Is that the correct interpretation?
Many thanks for your help so far.
Anna
Untitled Post
Posted on 07/30/10 11:01:58
Number of posts: 394
Hi Anna
To help you with your interpretation I need a bit more information. First, I assume you did it as a two group analysis. If so, then I need to know what the LV from the behavior side looked like. In other words, what was the correlation of brain scores with your behavior measure in each group?
Please keep in mind that a difference between groups in brain scores is still not surprising, even with the behavior PLS. Given GMV differences that already exist, this will never be a surprise. What the key interest for you is the correlation between brain and behavior and whether it differs. This comes from the behavior PLS analysis directly.
Also in your email you wrote:
To help you with your interpretation I need a bit more information. First, I assume you did it as a two group analysis. If so, then I need to know what the LV from the behavior side looked like. In other words, what was the correlation of brain scores with your behavior measure in each group?
Please keep in mind that a difference between groups in brain scores is still not surprising, even with the behavior PLS. Given GMV differences that already exist, this will never be a surprise. What the key interest for you is the correlation between brain and behavior and whether it differs. This comes from the behavior PLS analysis directly.
Also in your email you wrote:
I chose Behav PLS to run and entered the behav score in the same order as the images were uploaded; patients stacked on top of controls.
I assume you got this backwards as its the opposite order as for the images.
Randy
Untitled Post
Posted on 08/02/10 02:56:57
Number of posts: 18
A two group analysis? I think this may be where I have made a fundamental mistake. How does one enter a 2 group analysis in the GUI?
Intuitively I would have thought that when loading in my behav.txt file I should have had one column for group membership and then another column for behavioural score and then one LV would load on group diff and the other on behaviour........
And why does it matter what order the images are in as long as the behaviour files are in the same order as the images?
The brain scores plotted against the behave have a strong negative correlation although I haven't plotted it out to see if the groups are clustered at either end of the scale (they probably are to some degree if I'm getting a mean diff).
Thank you for your patience :)
Anna
Untitled Post
Posted on 08/02/10 09:27:51
Number of posts: 394
A two group analysis? I think this may be where I have made a fundamental mistake. How does one enter a 2 group analysis in the GUI?
When you set up the analysis, you add the second group with the 'add' button. The behavior file stays the same. You will need to construct a separate data matrix for each group (sorry that I was not clear on this before).
Intuitively I would have thought that when loading in my behav.txt file I should have had one column for group membership and then another column for behavioural score and then one LV would load on group diff and the other on behaviour........
Actually you would need a third column for the crossproduct of the two vectors to code for group differences in the brain-behavior relationship, but this is not necessary for PLS.
And why does it matter what order the images are in as long as the behaviour files are in the same order as the images?
As long as both files are in the same order, the world is a safe place.
Let me know how the analysis goes.
Randy
Untitled Post
Posted on 08/02/10 11:39:51
Number of posts: 18
Ahh its all beginning to make sense now :) Thanks so much for taking the time
I have one more question.
If I did just want to find the LV image that differentiated the groups could I just use 1 group analysis with behav PLS and then a text file coded 1 and 0 for group membership?
Many many thanks
Anna
Untitled Post
Posted on 08/02/10 11:42:49
Number of posts: 394
You certainly could get the group differences that way. You can also use the mean-centred PLS where you would not need to code group membership.
Randy
Randy
Untitled Post
Posted on 08/16/10 11:12:16
Number of posts: 18
Hi Randy
You said in the previous post "actually you would need a 3rd column with the cross product of the vectors to account for the interactions of group with behaviour but this is not necessary for PLS" Do you mean it's optional or that there is another way to get at the interactions? e.g put the brain scores and behav measure into a stats package and testing for a group interaction? The reason I ask is that I did put in the cross product column and now I don't know how to 'read' the brainscore vs behav plots since of course it fits a line through all the data points as if they are in one group.
Thanks
Anna
Login to reply to this topic.
